Feature selection using forest optimization algorithm (FSFOA) has well classification performance and dimensional reduction ability, but it has variable quality of initialised forest, larger randomness of local seeding and global seeding, and the low computational efficiency caused by expensive fitness evaluation. To solve the above problems, this paper proposes a feature selection using greedy-like forest optimization algorithm based on scoring mechanism (FSGLFOA-SM). Firstly, a scoring mechanism is constructed by using the classification accuracy of each dimensional decision variable as its score. From this, a greedy-like initialization strategy is proposed to generate an initialised forest with better quality, and a greedy-like local seeding strategy is proposed based on the comparison of scores, so that the decision variables with relatively higher scores can get a larger probability of local seeding. Then, a greedy-like genetic operator seeding strategy is proposed in the global seeding stage. The candidate forest is obtained by optimal selection and reconstruction, on which genetic, greedy-like crossover and mutation are carried out. So that the feature dimensions with higher scores are more likely to be retained and the classification accuracy of global seeding stage can be improved. Finally, a historical database is established, which reduces the calculation of duplicate solutions through accessing the database before fitness evaluation. The experimental results show that FSGLFOA-SM has superior classification accuracy and dimension reduction on 16 UCI datasets compared to the nine feature selection algorithms.